Spring 2020

Content of this presentation

  • Introduction to the project
  • Project overview
  • Methods used
  • Results
  • Discussion
  • Conclusion

Introduction

  • Prediction of protein-protein interactions (PPI) are a challenging task.
  • Machine learning (ML) models allow to exploit the content of PPI data sets generated by deep mutational scanning.
  • The aim of this project is to create a toolbox for generating structure-function insights using visualization tools and subsequently allow for predictions the biological activity of proteins and peptides using ML models.
  • Features of the toolbox:
    • Input and output data Visualization options.
    • Support for both sequence or variant input.
    • Support for several sequence encoders.
    • Support for several models.

Project workflow

Project overview

Deep mutational scanning

  • Deep mutational scanning allows for assessing the functional consequences of up to hundreds of thousands of variants of a protein in a single experiment.

  • It combines high-throughput DNA sequencing with a selection in which a physical association is maintained between each protein variant and the DNA that encodes it.

  • The sequence analysis provides the frequency of each variant in an input population and in a population after selection, with this ratio serving as a proxy for the function of each variant.

Methods - the data sets

Protein Target Biological activity Species Num of variants Score
Data set 1 BRCA1 BARD1 RING domain Ubiquitin E3 activity H. sapiens 5610 Y2H assays
Data set 2 ERK2 Small molecule (SCH772984) Resistance to drugs H. sapiens 6810 Drug sensitivity assays. Calculation of cell availability
Data set 3 LDLRAP1 OBFC1 Protein translation H. sapiens 6385 Y2H assays
Data set 4 Pab1 el4FG1 Translation initiation S. cereviseae 1340 Y2H assays

This presentation

  • In this presentation we focus on a particular dataset; data set 4 concerning deep mutatagenesis of Pab1.
  • We did however run all four data sets through the pipeline to verify the generality of the workflow and to see if some amino acid decriptors are better than others depending on the protein function.

Methods - packages used

Function Library
Data loading readxl
Data cleaning and wrangling dplyr , broom (tidyverse)
Data augmenting dplyr (tidyverse),Peptides
Extracting data UniprotR
Plotting ggplot2(tidyverse), ggseqlogo,ggpubr
Analysing stats
Modeling keras,neuralnet, caret, yardstick, glmnet,ANN2

Score densities for Pab1 RRM2 domain variants

Residues allowed in Pabl1 RRM2 domain

An amino acid was counted as active if score >0.

Heatmap of Pabl1 RRM2 domain variant scores

The scoring scale was truncated to between -2…2.

Sequence logo of Pabl1 RRM2 domain(135-185) variant scores

The score was normalized and summed to 1 for each postion.

Machine learning toolbox

  • Supported machine learning framework:

    • Artificial Neutral Network (ANN) regression.
    • Regularized linear regression.

Amino acid encoding

  • In order to present the protein sequences to the ML framework we used amino acids descriptors.
  • Each amino acid residue is converted to a vector and each protein as concatenation of vectors to yield a highly dimentional vector representation.
  • The following amino acid scales are available:
Scale(s) Description of scale
blosum45,50,62,80,90 Substitution matrix based on VARIMAX analysis of physicochemical properties
pam30,70,250 Substitution matrix based on observed mutations in phylogenetic trees
z5_scales PCA of physicochemical properties

Principal component analysis (PCA) showed around 200 components needed

Artificial Neural Network with Keras and linear model

Cross validation of a Pabl1 RRM2 glmnet model with Z5-scale

Prediction of unknown Pabl1 RRM2 variants with glmnet and Z5-scales

Discussion and model comparison for all four data sets 1

  • For all four data sets we compared ANN and glmnet regression (ElasticNet) on two amino acid descriptor scales; Z5-scales and blosum62.
  • We found the model performance (meassured as RMSE) for the test sets to be comparable across both scales and regressor - except for data set 2 where ANN seemed to perform better than glmnet.

Discussion and model comparison for all four data sets 2

  • The ANN with keras was performed for all 4 data set obtaining a really low accuracy for all of them. It was necessary to downsize the datasets, taking 1/4 of the sequence, in order to obtain a final model in each case.
  • The models performance is not impressive. Most likely other model types such as gaussian process regression or more advanced neural networks might be needed.
  • None the less we were able to predict the missing variant for data set 4 using the glmnet regressor which we implemented in a shiny app.

Shiny App

Challenges

  • RStudio Cloud is limited in memory! We had a lot of crashes both with keras, ANN2 and glmnet in the cloud.
  • All modelling and predications of full scale data set are possible using a local RStudio installation using the idendicated libraries.
  • On the other hand we had problem with local keras installations hence the need for ANN2 and glmnet.

Conclusion

  • Visualization of the raw data provided insights into the SAR of the 4 proteins.
  • We are able to find regions in the proteins which are both sensitive and insensitive to substitutions.
  • The heatmap shows quite that a substantial part of the C-terminal region studied was not very well covered in the scanning.
  • Using machine learning based on amino acid descriptors we were able to generate predictive models

References

  • Data set 1: L. M. Starita, D. L. Young, et al. Massively parallel functional analysis of brca1 ring domainvariants, Genetics, vol. 200, no. 2, pp. 413–422, 2015.
  • Data set 2: L. Brenan, A. Andreev, et al. Phenotypic characterization of a comprehensive set of missense mutants. Cell Reports, vol. 17, no. 4, pp. 1171–1183, 2016.
  • Data set 3: A deep mutational scan of LDLRAP1 based on a Y2H assay with the interactor OBFC1. [https://www.mavedb.org/scoreset/urn:mavedb:00000036-a-1/]
  • Data set 4: D. Melamed, D. L. Young, et al. Combining natural sequence variation withhigh throughput mutational data to reveal protein interaction sites. PLOS Genetics, vol. 11, no. 2,pp. 1–21, 2015.

References

Score density collection

Heatmap collection

Sequence logo collection

Residues allowed collection

Appendix

R script overview 1

  • 02_clean.R
    • Load: Load data from 01_load.R
    • Wrangle data: Remove NaN, fixes
    • Save cleansed data in .tsv format
  • 03_augment.R
    • Load data from 02_clean.R
    • Augment data: Calculate sequences, descriptors (pI, MW, Z5-Scale)
    • Save augmented data in .tsv format

Appendix

R script overview 2

  • 04_visualization.R
    • Load augmented data
    • Makes various plots
  • 05_PCA.R
    • PCA of the dataset
    • Scree - biplotting

Appendix

R script overview 3

  • 06_modelling_ANN_keras.R,
  • 06_modelling_ANN2.R,
  • 06_glmnetCV.R,
  • 06_modelling_glmnet
    • Performs model cross-validation
    • Performs model fitting
    • Performs model quality calculation
    • Models: ANN, glmnet

Appendix

R script overview 4

  • 03_get_unknowns.R
  • Gets variant_IDs for NA values
  • 03_augment_unknowns.R
  • Augment with protein sequence
  • 07_predict.R
  • Predict unknowns and makes a plot.

Appendix

R script overview 5

  • 99_proj_func.R
    • Sequence encoder
    • Sequence generator
    • Plot functions